Suhail(@Suhail)

I would love a Stripe Press of all the essays. @patrickc

Suhail (@Suhail) — Sat, 03 Oct 2026 22:23:30 GMT

I would love a Stripe Press of all the essays. @patrickc

Paul Graham@paulg

I recently met someone who'd printed out all my essays. I'd never seen this before and I was surprised how big they were, like two encyclopedia volumes (for anyone who even knows what an encyclopedia volume is).

🔗 View Quoted Tweet

💬5🔄0❤️34👀3960📊6

⚡ Powered by xgo.ing

26/ new biggest day

Suhail (@Suhail) — Sat, 03 Oct 2026 18:42:32 GMT

26/ new biggest day

💬1🔄0❤️3👀1693📊1

⚡ Powered by xgo.ing

🎯 AI-mind virus

Suhail (@Suhail) — Sat, 03 Oct 2026 15:01:47 GMT

🎯 AI-mind virus

Sam Altman@sama

I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue.

🔗 View Quoted Tweet

💬9🔄2❤️99👀24301📊14

⚡ Powered by xgo.ing

25/ biggest revenue day we've had thus far

Suhail (@Suhail) — Fri, 02 Oct 2026 18:04:59 GMT

25/ biggest revenue day we've had thus far

💬2🔄1❤️8👀2520📊3

⚡ Powered by xgo.ing

Congrats @natolambert - excited for you. In the same way that I believe intelligence should be flatt...

Suhail (@Suhail) — Fri, 02 Oct 2026 16:45:10 GMT

Congrats @natolambert - excited for you. In the same way that I believe intelligence should be flattened and diffused broadly, the knowledge to build this technology should be as well. I can't think of anyone better!

Nathan Lambert@natolambert

Today we're unveiling Trillium Labs @trillium_labs, a new non-profit to foster the open science of frontier AI. We're building open post-training recipes and will expand into open infra to study RSI, reward-hacking, multi-agent systems, and whatever comes next.

We're built around the theory of change that you need more eyes to solve hard technical problems. We have faith in the scientific methods and communities that humanity has built, and worry that AI is becoming too closed to utilize them.

Trilliums are wildflowers that bloom briefly in the spring, before the forest canopies fill out. Though they are small, they lay the foundation for the cycles of growth and nourishment through the rest of the year. At Trillium Labs, the recipes will be the slow nutrients for the seasons and the model releases will be the blooms. Building an institution dedicated to this is needed because, much as nature’s trilliums are slow to expand and grow, the open-ecosystem needs time and dedicated resources to catch up.

I co-founded with with a long-time friend and collaborator Tom Zick (@thesezickbeats). We're hiring (full time + student collabs/interns), we're fundraising, and we're looking for compute. Please get in touch if you're interested in helping out. Offices based in the Bay Area and Cambridge MA, remote okay.

I’m in the Bay Area until for The Curve and COLM to connect with people who are interested. We’re thankful to have initial support from Halcyon Futures and Schmidt Sciences with more funding en route to enable our ambitions of scaling. Our advisors @Thom_Wolf, @HannaHajishirzi, @gneubig and @ctnzr have been instrumental to building the ecosystem that exists today, and I’m stoked to get to keep working with them.

🔗 View Quoted Tweet

💬4🔄4❤️43👀14674📊9

⚡ Powered by xgo.ing

The Vera Rubin era just booted up. 🫡

Suhail (@Suhail) — Fri, 02 Oct 2026 16:36:54 GMT

The Vera Rubin era just booted up. 🫡

💬14🔄28❤️729👀71186📊99

⚡ Powered by xgo.ing

Dang, this dot thing really built Mighty into Codex. I’ve come full circle. The great cloud compute...

Suhail (@Suhail) — Thu, 01 Oct 2026 20:41:50 GMT

Dang, this dot thing really built Mighty into Codex. I’ve come full circle.

The great cloud computer revolution has finally begun.

💬25🔄8❤️142👀21132📊40

⚡ Powered by xgo.ing

Get your .si domain people! Appreciate @rob at @snagged for helping us lock it down :)

Suhail (@Suhail) — Wed, 30 Sep 2026 21:41:04 GMT

Get your .si domain people! Appreciate @rob at @snagged for helping us lock it down :)

💬35🔄13❤️286👀85925📊82

⚡ Powered by xgo.ing

24/ it was very important to me this time around that progress in ai (from anywhere) would lead to m...

Suhail (@Suhail) — Wed, 30 Sep 2026 01:06:42 GMT

24/ it was very important to me this time around that progress in ai (from anywhere) would lead to mostly good days vs tough days (ie you don’t get disrupted) and today was a good test of that still being true

💬1🔄0❤️5👀2622📊2

⚡ Powered by xgo.ing

🎯

Suhail (@Suhail) — Wed, 30 Sep 2026 00:37:10 GMT

🎯

dax@thdxr

i actually do like superintelligence a lot more than ai

🔗 View Quoted Tweet

💬2🔄0❤️45👀15832📊5

⚡ Powered by xgo.ing

🎯

Suhail (@Suhail) — Tue, 29 Sep 2026 21:57:58 GMT

🎯

will brown@willcb

openai announces glm-5.3 in codex subscriptions

anthropic announces glm-5.3 is evil and chinese

🔗 View Quoted Tweet

💬1🔄4❤️94👀26472📊9

⚡ Powered by xgo.ing

I think a lot of the fear that AI has on the world can be replaced with thinking more ambitiously. I...

Suhail (@Suhail) — Tue, 29 Sep 2026 14:43:16 GMT

I think a lot of the fear that AI has on the world can be replaced with thinking more ambitiously. Instead of being a software engineer, you can think of yourself as a Creator of Open Worlds. It will still be advantageous to be technical but also creative simultaneously.

Aravind Srinivas@AravSrinivas

Perplexity Computer (in Standard Mode with Opus 5.5) is capable of designing and building an end-to-end, browser-based open-world game by renting its own GPUs, running headless Blender, doing procedural design, downloading assets online, and designing audio. Done with ~$2000 worth of credits. Play here: vice.pplx.app

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬21🔄15❤️182👀40092📊38

⚡ Powered by xgo.ing

Three thoughts after playing with Jev: - this is way more than a simplistic classifier: you can pac...

Suhail (@Suhail) — Tue, 29 Sep 2026 14:05:12 GMT

Three thoughts after playing with Jev:

- this is way more than a simplistic classifier: you can pack a lot of complex state and do more than you’d think - it’ll take a minute to think about your problems in a non-LLM shape

- not quite beating frontier LLMs in correctness most of the time so can’t quite switch yet but it’ll clearly get there

- this is definitely a new model type we will all use by next year; I wouldn’t ignore it

💬40🔄24❤️545👀52013📊118

⚡ Powered by xgo.ing

We are hurtling very quickly towards everyone being capable of doing great AI research/eng. It is mo...

Suhail (@Suhail) — Sun, 27 Sep 2026 14:40:54 GMT

We are hurtling very quickly towards everyone being capable of doing great AI research/eng. It is mostly a game of who can accumulate enough compute and power to build/scale their ideas.

💬26🔄5❤️140👀14951📊39

⚡ Powered by xgo.ing

🎯

Suhail (@Suhail) — Sun, 27 Sep 2026 14:25:49 GMT

🎯

💬12🔄30❤️942👀74934📊132

⚡ Powered by xgo.ing

23/ brand is art

Suhail (@Suhail) — Fri, 25 Sep 2026 22:24:27 GMT

23/ brand is art

💬1🔄0❤️7👀2839📊2

⚡ Powered by xgo.ing

🎯

Suhail (@Suhail) — Fri, 25 Sep 2026 15:05:13 GMT

🎯

Jawwwn@jawwwn_

Mark Zuckerberg: failure doesn’t matter.

“The question is how much you win, when you win.”

“@WalterIsaacson wrote this biography of Einstein. One of the interesting things were all the theories he had that were wrong.”

“People don’t remember those.”

“We’re all going to get a lot of stuff wrong, what matters are the intellectual and cultural contributions we make that help push stuff forward.”

“We have a resilient company that is not afraid of failure.”

“That allows us to take some big swings and tolerate being ridiculed for long periods of time.”

“If you’re trying to do something that’s really new, people have to think it’s ridiculous. If they didn’t, someone else would’ve already done it.”

“In creation, the upside is infinite.”

“If you build something that hundreds of millions, or billions of people like, then that makes up for a lot of ideas that didn’t end up working.”

Via @Complex

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬5🔄40❤️640👀79962📊88

⚡ Powered by xgo.ing

What are good tools for LLM data inspection? Looking at SFT traces, RL env tasks, etc.

Suhail (@Suhail) — Thu, 24 Sep 2026 16:09:03 GMT

What are good tools for LLM data inspection? Looking at SFT traces, RL env tasks, etc.

💬19🔄5❤️98👀16603📊42

⚡ Powered by xgo.ing

The Slack bot is especially good. Have been pleasantly surprised.

Suhail (@Suhail) — Thu, 24 Sep 2026 03:22:02 GMT

The Slack bot is especially good. Have been pleasantly surprised.

💬0🔄0❤️12👀2953📊2

⚡ Powered by xgo.ing

Linear is so good. I hadn’t really gotten to use it until recently. Lovely UX end to end. I can te...

Suhail (@Suhail) — Thu, 24 Sep 2026 01:26:19 GMT

Linear is so good. I hadn’t really gotten to use it until recently. Lovely UX end to end. 

I can tell the team truly uses the product and fixes all the little paper cuts.

💬43🔄13❤️335👀29568📊74

⚡ Powered by xgo.ing

Looking for an office that can fit 6-8 people in San Francisco. DM me if you have extra space or y...

Suhail (@Suhail) — Tue, 22 Sep 2026 22:57:59 GMT

Looking for an office that can fit 6-8 people in San Francisco. 

DM me if you have extra space or you're looking for a friendly to take the space.

💬7🔄0❤️22👀14790📊9

⚡ Powered by xgo.ing

Super cool!

Suhail (@Suhail) — Tue, 22 Sep 2026 21:48:53 GMT

Super cool!

Perplexity@perplexity_ai

New research: We post-trained a Computer model to learn from its own errors using hint-guided self-distillation.

In a live A/B test, a later trained checkpoint reduced tool-call failures by 21.2% relative to an earlier checkpoint.

🔗 View Quoted Tweet

💬5🔄1❤️37👀14640📊9

⚡ Powered by xgo.ing

This is not a drill. 10x stronger than fentanyl. I am just sharing it widely because I find this ...

Suhail (@Suhail) — Sun, 20 Sep 2026 17:51:05 GMT

This is not a drill. *10x stronger* than fentanyl.

I am just sharing it widely because I find this kind of thing very concerning and it's getting almost no attention.

Brooke Jenkins 謝安宜@BrookeJenkinsSF

A new synthetic opioid, cychlorphine, up to 10x stronger than fentanyl, has reached SF— and it doesn't register on fentanyl test strips. It has already taken the life of a 16 year old San Franciscan.

Your browser does not support the video tag.

🔗 View on Twitter

🔗 View Quoted Tweet

💬32🔄25❤️504👀134596📊74

⚡ Powered by xgo.ing

22/ 500B+/week

Suhail (@Suhail) — Sun, 20 Sep 2026 16:12:23 GMT

22/ 500B+/week

💬1🔄0❤️1👀3074📊1

⚡ Powered by xgo.ing

🎯

Suhail (@Suhail) — Sun, 20 Sep 2026 13:30:35 GMT

🎯

Nathan Lambert@natolambert

My best guess is that for scaled RL most of the top Chinese AI labs are starting to use a lot of Huawei for inference and Nvidia for training (maybe not for weird architectures). As agent swarms, even more scaled post-training, etc becomes the norm, this will accelerate their domestic industry.

🔗 View Quoted Tweet

💬1🔄0❤️37👀20215📊5

⚡ Powered by xgo.ing

Yep. It’s so early.

Suhail (@Suhail) — Sat, 19 Sep 2026 04:01:42 GMT

Yep. It’s so early.

Ayush Gundawar@gundawar

This slide from @flappyairplanes’s talk at Sequoia’s AI Ascent is pretty cool.

For those working inside any one circle, it generally feels fairly obvious that the application of AI to that industry has a large TAM. But the birds-eye view is still striking -- we are quite early in the diffusion process, and many high-TAM surfaces remain ‘unsolved’

🔗 View Quoted Tweet

💬9🔄9❤️199👀59858📊48

⚡ Powered by xgo.ing

After that it’s offload, offload, offload.

Suhail (@Suhail) — Sat, 19 Sep 2026 03:50:27 GMT

After that it’s offload, offload, offload.

💬0🔄0❤️4👀2611📊1

⚡ Powered by xgo.ing

Engram tables in DSV4.1 are super cool! Everything is just compression, compression, compression. Af...

Suhail (@Suhail) — Sat, 19 Sep 2026 02:36:03 GMT

Engram tables in DSV4.1 are super cool! Everything is just compression, compression, compression. After that, it's all caching, caching, caching.

💬11🔄1❤️38👀14996📊15

⚡ Powered by xgo.ing

To be a little positive, SV of 2026 is a generation out of distribution to 2001-2010. While the roun...

Suhail (@Suhail) — Thu, 17 Sep 2026 01:51:35 GMT

To be a little positive, SV of 2026 is a generation out of distribution to 2001-2010. While the rounds are enormous here, it was the norm to invest in growth without revenue. B2B hardly existed.

OTOH, SV traditionally roots against unjustified large rounds and has typically resulted in  startups failing. Usually you want the tech world on your side as the underdog.

robin@zebird0

Instinct $10B round giving 2021 Clubhouse vibes right now 

- Great usage and growth 
- No monetization yet
- Fast follow competition from incumbents

Will be interested to see how this evolves

🔗 View Quoted Tweet

💬3🔄0❤️11👀2646📊3

⚡ Powered by xgo.ing

Wild to watch to see prices of these GPUs going up still

Suhail (@Suhail) — Wed, 16 Sep 2026 22:47:44 GMT

Wild to watch to see prices of these GPUs going up *still*

💬1🔄2❤️28👀7068📊5

⚡ Powered by xgo.ing

21/ Always check your NICs. Currently in broken NIC firmware hell.

Suhail (@Suhail) — Wed, 16 Sep 2026 22:22:52 GMT

21/ Always check your NICs. Currently in broken NIC firmware hell.

💬1🔄0❤️0👀3065📊1

⚡ Powered by xgo.ing

--- name: deslop-shared-libs description: Find places where shared code could reduce duplication and...

Suhail (@Suhail) — Wed, 16 Sep 2026 18:25:16 GMT

---
name: deslop-shared-libs
description: Find places where shared code could reduce duplication and improve reliability, starting with the past two weeks of commits and PRs. Use for /deslop-shared-libs, $deslop-shared-libs, or requests to find code worth sharing. Give recommendations without changing code.
---

# Find code worth sharing

Find five opportunities to share code, then recommend the best three. Focus on
areas people are actively changing and changes that would delete more code than
they add.

## Scope

Review only. Do not edit code, install packages, open issues or PRs, or deploy.
Follow any narrower scope the user gives.

## Start with recent work

- Check the default branch's latest commit and choose a review commit. Review
  the past 14 days plus relevant current-branch work. State the commit and dates,
  and check candidate files against it if the checkout differs. Note any limits
  on access or history. Use a different time window if requested.
- Read PRs opened, updated, or merged in that window using tools such as `gh`.
  Check the changed files and relevant diffs. Separate code changes from comment
  and bot noise, and count a merged PR and its commits as one effort.
- Start where you see repeated fixes or similar code being added. Follow related
  callers and shared helpers, including older files. Frequent changes alone do
  not make something worth sharing.
- Check whether PR code has merged and whether someone is already doing the
  proposed work. Verify findings against the source you reviewed. If history or
  PR access is missing, say so and use what is available; missing data does not
  mean an area is inactive.

## Look for useful shared code

- Check existing libraries and helpers first. Reuse proven code where possible.
- Cover the active areas in each language before ranking. Compare functions,
  types, schemas, and configuration, including copies with different names or
  formatting. Check the related files around each strong find for more copies.
- For each of five ideas, show at least two places that could share code. Link
  to files, functions, and lines. Check that they need the same behavior;
  looking similar is not enough.
- Describe the smallest useful helper and where it would live. Check whether
  callers can share it across languages, dependencies, and deployment setups.
  Preserve differences that callers need.
- Explain how sharing would improve reliability. Point to tests or established
  use, and note missing coverage or how a shared bug could affect more callers.
- Base estimates on named code blocks: what goes away, what replaces it, and
  what tests and integration work are needed. Show implementation savings and
  total savings separately; favor total reductions. Use ranges when uncertain.
  Exclude generated and third-party code; count moved code on both sides.

## Pick the best three

- Set aside work already covered by an open or merged PR. Mention it separately;
  it does not count toward the five new ideas.
- Favor code savings and reliability gains. Prefer recently changed areas when
  benefits are similar; explain any choice to focus on older code.
- Favor changes that are easy to adopt and test, with a low risk of changing
  behavior. Prefer code already tested and used in several places. Avoid helpers
  with too many options or ones that tie unrelated components together.
- Choose three and explain why the other two rank lower. Report fewer if the
  evidence does not support five ideas or three worthwhile recommendations.
  Say when a change might add more code than it removes.

## Output

Keep explanations short and plain spoken.

1. Briefly state the dates, sources, and areas reviewed, including any gaps.
2. Compare up to five new ideas in a compact table: affected code, commit or PR
   links, estimated total lines removed, added, and saved, reliability benefit,
   and whether each made the top three.
3. Rank the best three. Link to the code, describe what to share, explain why it
   is useful now, and name the main risk or uncertainty. Briefly explain why the
   other two ranked lower and how recent activity affected the choice.

Label estimates clearly. Stop after the recommendations.

💬0🔄0❤️15👀3038📊7

⚡ Powered by xgo.ing

New skill I made: /deslop-shared-libs Find places where shared code could reduce duplication and im...

Suhail (@Suhail) — Wed, 16 Sep 2026 18:25:09 GMT

New skill I made: /deslop-shared-libs

Find places where shared code could reduce duplication and improve reliability, starting with the past two weeks of commits and PRs.

Lately, I find that AI agents love to write lots of net-new code. They often don't intertwine things, modularize systems, or make shared components. This adds a ton of tech debt into the system. It makes it write more tests. Slows compile times.

Sometimes I am just idling waiting for something to finish so this is a sorta new code golf where I can routinely de-slop the code base.

💬17🔄8❤️93👀15286📊40

⚡ Powered by xgo.ing

20/ step 1: just make a useful thing step 2: just make it better step 3: just make it better again (...

Suhail (@Suhail) — Wed, 16 Sep 2026 05:23:59 GMT

20/
step 1: just make a useful thing
step 2: just make it better
step 3: just make it better again (x100 days)
step 4: profit

💬2🔄0❤️5👀3383📊3

⚡ Powered by xgo.ing

@martin_casado I think one way to imagine things differently is if you (1) didn't worry about the co...

Suhail (@Suhail) — Wed, 16 Sep 2026 00:50:02 GMT

@martin_casado I think one way to imagine things differently is if you (1) didn't worry about the cost of inference that much and (2) it was near-instant.

💬2🔄2❤️25👀9261📊6

⚡ Powered by xgo.ing

If you do > 1B tokens a day in batched api requests with OpenAI or Anthropic, would love to talk ...

Suhail (@Suhail) — Tue, 15 Sep 2026 23:09:55 GMT

If you do > 1B tokens a day in batched api requests with OpenAI or Anthropic, would love to talk to you to learn about ups/downs of implementing batching. 

DM

💬16🔄2❤️65👀16360📊21

⚡ Powered by xgo.ing

This will happen to Inference

Suhail (@Suhail) — Tue, 15 Sep 2026 03:04:58 GMT

This will happen to Inference

Paul Graham@paulg

One data point in technology making everyone richer: the cost of lighting has decreased by over 1000x.

🔗 View Quoted Tweet

💬29🔄18❤️372👀80116📊57

⚡ Powered by xgo.ing

19/ we are finally, somewhat reliable ❤️

Suhail (@Suhail) — Tue, 15 Sep 2026 02:24:33 GMT

19/ we are finally, somewhat reliable ❤️

💬3🔄0❤️8👀3215📊3

⚡ Powered by xgo.ing

18/ critical third hire made 🎉

Suhail (@Suhail) — Fri, 11 Sep 2026 22:24:49 GMT

18/ critical third hire made 🎉

💬1🔄1❤️3👀3139📊2

⚡ Powered by xgo.ing

There's nothing that I hate more these days than this moment.

Suhail (@Suhail) — Fri, 11 Sep 2026 03:30:01 GMT

There's nothing that I hate more these days than this moment.

💬5🔄2❤️21👀2431📊7

⚡ Powered by xgo.ing

Out of curiosity, how does one build something that coordinates 10K AI agents to solve a single prob...

Suhail (@Suhail) — Tue, 08 Sep 2026 19:05:48 GMT

Out of curiosity, how does one build something that coordinates 10K AI agents to solve a single problem?

💬327🔄39❤️1641👀297254📊485

⚡ Powered by xgo.ing

--- name: cleanbranch description: Create a new branch with only the essential feature changes. Use ...

Suhail (@Suhail) — Sun, 06 Sep 2026 17:32:47 GMT

---
name: cleanbranch
description: Create a new branch with only the essential feature changes. Use for /cleanbranch, $cleanbranch, or requests to extract a minimal feature branch.
---

# Clean branch

Create the smallest complete PR for the intended feature, with clear code and minimal comments.

- Infer the feature from the conversation and review existing changes, including committed,
  uncommitted, and untracked work.
- Create a fresh branch from the user's chosen base, the current PR's base, or the repository's
  default branch. Use a descriptive branch name unless one is provided. Prefer a separate worktree
  to preserve the original branch and working files.
- Carry over only what the feature needs to work end to end, including necessary integration
  and tests. Leave out unrelated refactors, formatting changes, experiments, and speculative work.
- Favor clear names, simple control flow, and existing patterns. Avoid unnecessary abstractions.
  Comment only on non-obvious intent or constraints; remove comments that repeat the code.
- Re-verify that the feature works end to end, maintains comparable performance and quality,
  and satisfies the original requirements and intended scope.
- Review the full diff against the base and run relevant tests and repository-required checks.
  Fix failures introduced by the extraction and report any validation limitations.
- Report the branch, base, worktree path, what was kept or omitted, and validation results.
  Commit, push, or open a PR only when requested or already authorized.

💬2🔄1❤️32👀9599📊15

⚡ Powered by xgo.ing

How I go from 100K LOC to a few thousand these days: Lately I find that I enjoy having a very messy...

Suhail (@Suhail) — Sun, 06 Sep 2026 17:31:14 GMT

How I go from 100K LOC to a few thousand these days:

Lately I find that I enjoy having a very messy branch with lots of twists and turns, tests for verification, scripts to load test/adversarially break things, and experiments. Sometimes I am trying things to understand the pieces better. Rarely do I want all that bloat and slop to enter the main branch of the repo.

Please copy or tell me how to do it better.

💬20🔄6❤️306👀38516📊124

⚡ Powered by xgo.ing

@martin_casado Nah, not a skill issue. I always try one-shotting things with /goal now to see if it ...

Suhail (@Suhail) — Sun, 06 Sep 2026 17:20:13 GMT

@martin_casado Nah, not a skill issue. I always try one-shotting things with /goal now to see if it jumps in capability and I'd say I was fairly disappointed.

💬4🔄2❤️106👀26319📊16

⚡ Powered by xgo.ing

I hope it is clear that copying and following the herd for too long will unlikely lead to greatness ...

Suhail (@Suhail) — Sun, 06 Sep 2026 14:56:40 GMT

I hope it is clear that copying and following the herd for too long will unlikely lead to greatness but it’s a good way to get started while avoiding imagined markets with no real customers.

💬4🔄1❤️44👀4749📊9

⚡ Powered by xgo.ing

I would encourage more founders to copy the ideas out there that you’re passionate about. Not becaus...

Suhail (@Suhail) — Sun, 06 Sep 2026 14:49:21 GMT

I would encourage more founders to copy the ideas out there that you’re passionate about. Not because that is your end game, but to the degree it is something you enjoy building, you can get some customers and cockroach. That will give you time to adapt and figure out what unique thing it can evolve to as technology progresses.

In the age of AI, people can build very quickly. So, in many situations, it’s best to be in market with a team, funding, and some customers to take advantage of new capabilities in the models as they come out. It’s much harder to do what we used to do: quietly build a thing for 3-6 months and launch it. By the time you get the company started, you’ll have moved too slowly.

You should assume you will look no different than everyone adjacent to you in the first 3-4 months. However it may only be  like that for 3-4 months.

💬43🔄26❤️574👀42407📊145

⚡ Powered by xgo.ing

🎯 There are very few original startup ideas. It’s mostly adapting and learning from the market as ...

Suhail (@Suhail) — Sun, 06 Sep 2026 00:51:06 GMT

🎯

There are very few original startup ideas. It’s mostly adapting and learning from the market as you go. Most often you need to prepare for the moment especially these days.

Paul Graham@paulg

I talked to a founder who was demoralized because all the ideas he could think of could be easily duplicated. I told him to ignore that worry; nearly all startup ideas are like that at first. Their value is the other ideas they lead to.

🔗 View Quoted Tweet

💬32🔄33❤️679👀94505📊107

⚡ Powered by xgo.ing

Man, AI is insane.

Suhail (@Suhail) — Thu, 03 Sep 2026 19:51:35 GMT

Man, AI is insane.

💬39🔄26❤️590👀25332📊89

⚡ Powered by xgo.ing

Excited to see benchmarks headed in this direction and getting better!

Suhail (@Suhail) — Wed, 02 Sep 2026 19:58:31 GMT

Excited to see benchmarks headed in this direction and getting better!

Proximal@ProximalHQ

We are releasing FrontierSWE v2, our updated ultra-long horizon coding benchmark

V2 features an expanded task suite and improved methodology. We see large performance gaps between frontier models, with Claude Fable 5.1 leading by a wide margin

🔗 View Quoted Tweet

💬2🔄1❤️42👀15990📊7

⚡ Powered by xgo.ing

why is everything AI writes 10 billion lines of code?

Suhail (@Suhail) — Wed, 02 Sep 2026 03:35:11 GMT

why is everything AI writes 10 billion lines of code?

💬130🔄6❤️318👀45243📊137

Suhail(@Suhail)

I would love a Stripe Press of all the essays. @patrickc

26/ new biggest day

🎯 AI-mind virus

25/ biggest revenue day we've had thus far

Congrats @natolambert - excited for you. In the same way that I believe intelligence should be flatt...

The Vera Rubin era just booted up. 🫡

Dang, this dot thing really built Mighty into Codex. I’ve come full circle. The great cloud compute...

Get your .si domain people! Appreciate @rob at @snagged for helping us lock it down :)

24/ it was very important to me this time around that progress in ai (from anywhere) would lead to m...

🎯

🎯

I think a lot of the fear that AI has on the world can be replaced with thinking more ambitiously. I...

Three thoughts after playing with Jev: - this is way more than a simplistic classifier: you can pac...

We are hurtling very quickly towards everyone being capable of doing great AI research/eng. It is mo...

🎯

23/ brand is art

🎯

What are good tools for LLM data inspection? Looking at SFT traces, RL env tasks, etc.

The Slack bot is especially good. Have been pleasantly surprised.

Linear is so good. I hadn’t really gotten to use it until recently. Lovely UX end to end. I can te...

Looking for an office that can fit 6-8 people in San Francisco. DM me if you have extra space or y...

Super cool!

This is not a drill. *10x stronger* than fentanyl. I am just sharing it widely because I find this ...

22/ 500B+/week

🎯

Yep. It’s so early.

After that it’s offload, offload, offload.

Engram tables in DSV4.1 are super cool! Everything is just compression, compression, compression. Af...

To be a little positive, SV of 2026 is a generation out of distribution to 2001-2010. While the roun...

Wild to watch to see prices of these GPUs going up *still*

21/ Always check your NICs. Currently in broken NIC firmware hell.

--- name: deslop-shared-libs description: Find places where shared code could reduce duplication and...

New skill I made: /deslop-shared-libs Find places where shared code could reduce duplication and im...

20/ step 1: just make a useful thing step 2: just make it better step 3: just make it better again (...

@martin_casado I think one way to imagine things differently is if you (1) didn't worry about the co...

If you do > 1B tokens a day in batched api requests with OpenAI or Anthropic, would love to talk ...

This will happen to Inference

19/ we are finally, somewhat reliable ❤️

18/ critical third hire made 🎉

There's nothing that I hate more these days than this moment.

Out of curiosity, how does one build something that coordinates 10K AI agents to solve a single prob...

--- name: cleanbranch description: Create a new branch with only the essential feature changes. Use ...

How I go from 100K LOC to a few thousand these days: Lately I find that I enjoy having a very messy...

@martin_casado Nah, not a skill issue. I always try one-shotting things with /goal now to see if it ...

I hope it is clear that copying and following the herd for too long will unlikely lead to greatness ...

I would encourage more founders to copy the ideas out there that you’re passionate about. Not becaus...

🎯 There are very few original startup ideas. It’s mostly adapting and learning from the market as ...

Man, AI is insane.

Excited to see benchmarks headed in this direction and getting better!

why is everything AI writes 10 billion lines of code?

This is not a drill. 10x stronger than fentanyl. I am just sharing it widely because I find this ...

Wild to watch to see prices of these GPUs going up still